SHAP (SHapley Additive exPlanations)

SHAP (SHapley Additive exPlanations) is a model explanation method that attributes each prediction to the input features using Shapley values from cooperative game theory, so the contributions sum to the difference from the average prediction.

SHAP(섀플리 가법 설명)은 협력게임이론의 섀플리 값을 이용해 각 예측을 입력 특성별 기여도로 분해하며, 그 기여도의 합이 평균 예측과의 차이와 같아지는 모형 설명 방법입니다.

How do I explain which features drive the predictions of a machine learning model?

머신러닝 모형의 예측에 어떤 특성이 영향을 주는지 어떻게 설명하나요?

SHAP explains a trained model by assigning each feature a contribution to each individual prediction, and these local values can be aggregated into global feature importance and dependence plots. It fits when a predictive model, such as gradient boosting or random forest, needs to be interpreted, although the explanations describe the model's behavior rather than causal effects.

When to use SHAP (SHapley Additive exPlanations)

• Interpreting a gradient boosting, random forest, or neural network model
• Explaining why the model made a specific prediction for one case
• Showing the direction and size of each feature's effect on predictions

When to use something else

• Causal effects of a variable are the goal: use causal inference methods such as inverse probability weighting
• A quick model-agnostic global ranking is enough: use permutation feature importance

Assumptions and requirements

• A trained model with adequate predictive performance
• Representative background data for computing expectations
• Strongly correlated features make attributions harder to interpret
• Explanations describe the model, not causal mechanisms

What the analysis reports

• SHAP summary (beeswarm) plot
• Mean absolute SHAP values as global importance
• Dependence plots for individual features
• Waterfall or force plots for single predictions

How to report it

SHAP analysis indicated that [feature] was the most influential predictor, with a mean absolute SHAP value of [value], and higher values of [feature] increased the predicted [outcome].

Related methods

• Permutation Feature Importance
• LIME
• Partial Dependence Plot

Run SHAP (SHapley Additive exPlanations) on Quorant